Beijing Humanoid Robot Innovation Center unveils new patent that can enable robots to have zero-sample self-reflective capabilities.

The Qichacha APP shows that recently, Beijing Humanoid Robot Innovation Center Co., Ltd. has applied for the patent disclosure of "Robot Zero-Sample Self-Reflection Task Planning Method and System". The patent abstract shows that this disclosure relates to a method and system for robot zero-sample self-reflection task planning, including: analyzing scene color images through a zero-sample perception module to obtain scene feature descriptions, object detection boxes, and pixel-level mask information; inputting the scene feature descriptions and task instructions into the task planning model to generate a preliminary task plan, which is then further optimized by a reflection judgment model to generate a high-level task plan; determining the target object detection box based on the current sub-task, predicting its position and identifying its motion status through target tracking model; when the target object is stationary, generating candidate grasp poses by the grasp detection model based on its color depth image, generating grasp suggestion areas by the visual language model, and selecting the optimal grasp pose through evaluation metrics; after executing the sub-task, analyzing the task status by the reflection judgment model, continuing if successful, and performing analysis and planning corrections if unsuccessful, until the task is completed. This method enables robots to possess zero-sample self-reflection abilities, enhancing task robustness.